汉语和英语音高重音自动标注方法的对比与分析  

The comparison and analysis between Chinese Putonghua and English stress automatic annotation

作　　者：倪崇嘉[1,2] 刘文举 徐波 

机构地区：[1]山东财政学院统计与数理学院,济南250014 [2]中科院自动化研究所模式识别国家重点实验室,北京100190

出　　处：《声学学报》2012年第5期553-560,共8页Acta Acustica

基　　金：国家自然科学基金(90820303;60675026;90820011);国家高技术研究863计划(2006010124073;2006AA012194);国家重点基础研究发展计划973计划(2004CB318105);山东省自然科学基金(ZR2011FQ024)资助项目

摘　　要：虽然汉语和英语的重音自动标注被广泛的研究,但是关于汉语和英语的重音自动标注之间对比的研究还鲜有报道。基于汉语韵律标注库ASCCD和英语韵律标注库Boston University Radio News Corpus,对汉语和英语的重音自动标注的异同进行对比,考察不同的特征在不同语言的语料库上的泛化性能。通过基于集成分类回归树的重音自动标注实验、特征分析及基于互信息的重音自动标注的声学对比,得到如下结论:在相同的条件下,汉语重音自动标注的正确率比英语重音自动标注的正确率要低;在重音自动标注中,词典语法相关特征比声学相关的特征更重要;不同的声学信息源在重音自动标注中所起的作用不同,时长相关的特征对汉语和英语重音自动标注都很重要;英语中大部分特征提供的互信息要比汉语相应的特征提供的互信息要高。Although Chinese Putonghua stress automatic annotation and English pitch accent automatic annotation have been studied extensively, there is seldom works on exploring the comparison between Putonghua stress automatic annotation and English pitch accent automatic annotation. In this paper, we make comparison between Putonghua stress automatic annotation and English pitch accent automatic annotation based on Putonghua prosodic annotation corpus--ASCCD and English prosodic annotation corpus--Boston University Radio News Corpus, and examine the generalization of different features in different corpora. Through the experiments of stress automatic annotation based on boosting classification and regression tree, feature analysis and the acoustic comparison of stress automatic annotation based on the mutual information~ we get the following conclusions： （1） The correct rate in Putonghua stress automatic annotation is lower than that of English pitch accent automatic annotation; （2） When annotating stress, the lexical and syntactic related features are more important than the acoustic related features; （3） The different acoustic features have different function on stress annotation, and in all these features, the duration related features are important to Putonghua and English stress automatic annotation; （4） The mutual information value of most features in English is higher than that of in Putonghua.

关 键 词：自动标注方法 重音 英语 汉语 音高 BOSTON 特征比 分类回归树 

分 类 号：TN912.3[电子电信—通信与信息系统]